Specialist AI: When the General Model Is Not Enough
When the General Model Is Not Enough
The ten guides beside this one cover generalists. This one covers the specialists: purpose-built tools that beat a horizontal assistant inside one domain—coding, legal, finance, customer service, enterprise knowledge, healthcare. The recurring question is not “which is smarter,” it is “does this job clear the bar where a specialist’s data, workflow and guardrails are worth the price, the lock-in and the integration cost?”
A different question from “which model is best”
General assistants are extraordinary generalists. But in regulated, data-heavy or workflow-bound work, a tool that embeds the domain—the case law, the codebase, the EHR, the data room—often beats a smarter but generic model. The 2026 market has a credible specialist for most professional functions. The skill is knowing when to reach for one.
Domain layer over a frontier model
Most vertical tools run on the same OpenAI or Anthropic models as your general assistant. What you pay for is the layer on top: proprietary data, workflow, integrations, evaluation and accountability.
- The intelligence is often rented
- The domain layer is the product
- Judge the layer, not the demo
Depth a generalist cannot fake
A specialist embeds the things a general chat cannot: your repository, your matter history, your permission model, your compliance posture and the workflow the work actually follows.
- Context that lives in your systems
- Guardrails for regulated work
- Integrations into systems of record
Price, lock-in and overlap
Specialists cost more, embed your data deeply, and often overlap with the general assistant you already pay for. The decision is economic and architectural, not just about capability.
- Your Copilot may already do 80%
- Migration is rarely easy
- Buy only for the part that clears the bar
A specialist for most professional functions
This is the shape of the market in mid-2026: a well-funded, credible tool for most domains a general assistant handles only shallowly. Names change fast—treat this as a snapshot of categories, not a permanent leaderboard.
| Domain | What the specialist adds | Leading examples |
|---|---|---|
| Coding | Repository context, autonomous multi-file agents, CI and pull-request integration | Cursor (now owned by SpaceX), Cognition (Devin) |
| Legal | Case law, contracts, firm knowledge, citation discipline and conflict-of-interest enforcement | Harvey |
| Finance and research | Deep analysis across private documents, public filings and licensed market data | Hebbia |
| Customer service and long-horizon work | Outcome-priced agents across chat, voice, email and messaging, now extending to goals that run for weeks | Sierra |
| Enterprise knowledge | Permission-aware search and agents across all company apps, with agents that carry their own identity | Glean |
| Healthcare | Ambient clinical notes inside the EHR, extending into workflow and revenue cycle; evidence at the point of care | Abridge, OpenEvidence |
The most mature vertical, and the most contested
Software is where vertical AI is furthest along and where the general assistants also compete hardest. Claude Code and OpenAI Codex are the generalist coding agents; Cursor and Cognition are the purpose-built ones. The line between “feature of my assistant” and “dedicated tool” is blurriest here.
Cursor IDE agent · SpaceX-owned
Positions itself as a coding agent for building ambitious software: deep codebase understanding, autonomous and parallel agents, and bring-your-own-model across OpenAI, Anthropic, Gemini and others. Its enterprise page claims 64% of the Fortune 500 and more than 50,000 enterprises, with SOC 2 Type II, SAML and SCIM (vendor claims). The ownership changed on August 14, 2026: Cursor posted that it “has officially been acquired by SpaceX,” completing “the acquisition process that started in April, when we announced our partnership with SpaceXAI to accelerate our model training efforts.” No valuation and no terms are disclosed in the post, so do not put a number on it. Cursor Router, released July 22, adds model routing in Intelligence, Balance and Cost modes, which Cursor says delivers frontier-quality performance at 60% savings (vendor claim) — but the docs are explicit that “Cursor Router is currently only available on Teams and Enterprise plans,” so it is a policy lever for an organization, not something an individual developer can switch on. Cursor Start, announced July 28 at ₹649 per month for developers in India, tax inclusive, sits between Free and Pro. The August shipping list is long: Git at any scale (August 18), Origin Code Hosting in early beta (August 17), Firetiger joining Cursor, AIUC-1 certification and cloud agents starting 3x faster with builds (all August 13), Grok 4.6 support (August 12), a Router explainer (August 6) and Google Workspace Plugins (August 3).
- Acquired by SpaceX on August 14, 2026 — no terms disclosed
- Repo-aware, multi-model, shipping weekly
- Router is Teams and Enterprise only
- Ownership is now a procurement question, not a footnote
Cognition (Devin) Autonomous SWE
Operates Devin, marketed as the first autonomous software engineer: it plans, writes, tests and ships code inside your codebase and tools. Cognition ships named capabilities rather than version numbers — Devin Fusion, described on June 29, 2026 as a new kind of multi-model harness, and Devin Security Swarm; both were re-confirmed as the current framing on August 19, 2026, with no version number attached. On FedRAMP, read the exact words: “Cognition’s entire platform is now FedRAMP Class D (High) In-Process and listed on the FedRAMP Marketplace,” covering Devin Desktop (formerly Windsurf), Devin Cloud and the Devin CLI. In-Process is a place in a queue, not an authorization, and a public-sector buyer should treat it that way. A July 22 memorandum of understanding with the US Department of Energy joins the Genesis Mission across 17 national labs, alongside existing work with the US Army, Navy and NASA JPL. Cognition is also doing the buying: TierZero closed, and The Interaction Company of California now reads as absorbed rather than pending — the July 23 post says “Today, we’re welcoming The Interaction Company of California, the makers of Poke, to Cognition” and that “Poke users can continue using the product just as before.” An earlier edition of this guide called that deal announced but not closed; present-tense integration language is the stronger reading, though the post never uses the word “closed.” Cognition published nothing between August 3 and August 19, 2026 — the most recent post is the July 28 LTM partnership, which puts Devin into LTM’s financial-services client base and is distribution rather than a product change.
- Delegated, end-to-end tasks
- FedRAMP Class D (High) In-Process — a queue position, not an authorization
- TierZero closed; The Interaction Company now reads as absorbed
- Quiet window: no posts between August 3 and August 19, 2026
- Review every diff it ships
The generalist option Already yours
Claude Code and OpenAI Codex bring capable agentic coding inside tools you may already pay for. For many teams this covers the majority of the work before a dedicated tool is justified.
- No new vendor
- Strong for most tasks
- Benchmark against the specialists
Where domain data is the moat
In legal, finance and enterprise knowledge, the value is not raw intelligence—it is grounded access to the right documents, under the right permissions, with the right citation and audit discipline. These tools compete on the corpus and the controls, not the model. In July 2026 the controls moved fastest: identity for agents, ethical walls enforced inside the product, and licensed market data sitting beside your own files.
Harvey Legal
Agents built for law firms and enterprise legal teams, grounded in case law, contracts and firm knowledge. Reports more than 100,000 lawyers across 1,300 organizations, and raised at an $11B valuation in March 2026 — that figure still dates from March 25, 2026 and nothing since restates it. Read the July 28, 2026 strategic investment from Goldman Sachs and J.P. Morgan carefully for that reason: it publishes no new valuation, so do not infer one. Commercially, the July 16 acquisition of Y Combinator-backed Benchmark disclosed a record second quarter adding more than $100M of net-new annual recurring revenue, and 125-plus asset-management firms including Blue Owl, Bridgewater and KKR (vendor figures). The product news is August 18, 2026: Harvey II, announced with the line that “Harvey’s agents are smarter from the start. They begin with more than just what you put in the prompt.” There is no general-availability date and no pricing on that post — the call to action is a demo — so availability is unconfirmed and it does not belong in a dated rollout plan yet. Also in the window: Steve Zad appointed Chief Revenue Officer (August 4), HEUKING expanding Harvey firmwide (August 5), “The New Harvey for Outlook” (August 12) and a research post on training frontier review-table models with Applied Compute (August 14). AIUC-1 certification landed July 30.
- Harvey II announced August 18 — no GA date, no pricing, availability unconfirmed
- Goldman Sachs and J.P. Morgan invested July 28; no new valuation published
- The $11B valuation still dates from March 25, 2026
- A lawyer still signs the work
Hebbia Finance
Now describes itself as the leading AI platform for finance — narrower than the general research framing it once used. Its Matrix interface, with Skills and Agents, runs analysis across private documents plus public filings and licensed third-party market data: connectors include SEC filings, CapIQ, FactSet, PitchBook, Preqin, expert networks, Snowflake and SharePoint. A July 13 post on token efficiency reports more than 250B tokens a month (vendor figure). On July 30, 2026 Hebbia introduced Max, written up by Aabhas Sharma and described as “the first AI team member built for the way your firm works.” It is still not generally available: fetched again on August 19, 2026, that post still reads “Max is rolling out to a small set of firms first. Request a preview here.” Treat it as an invitation-gated preview, not something you can put in this quarter’s plan. One new post since, “Rethinking Control in the Age of AI Delivery” on August 4, 2026.
- Matrix, Skills and Agents
- Private, public and licensed data in one view
- Max is still a limited preview, re-confirmed August 19
- Verify every extracted figure
Glean Enterprise knowledge
Work AI that unifies permission-aware search, an assistant and agents across 100-plus company apps, so answers respect who is allowed to see what. Sits between a general assistant and a system of record. On July 17 Databricks querying arrived inside Glean Assistant through an MCP integration with Databricks Genie, with Databricks permissions, governance and security left intact (vendor statement). Three further changes are dated July 29, 2026, but they come from Glean’s release notes rather than a dated article — nothing in that window rendered on the Glean blog — so this guide states them without a link and a reader relying on any one of them should confirm it with Glean directly. First: Claude Opus 5, described as Anthropic’s latest premium Opus-class model, is now available in Glean Assistant model choice and in Agents. Second: Gemini 3.6 Flash is now available in Glean Assistant model choice and in the Agents model hub. Both require admin enablement and Glean notes they may be subject to usage-based pricing. Third: Glean is now available through the ChatGPT app marketplace, making it easier for admins to roll out Glean in ChatGPT.
- Search across all your tools
- Permission-aware by design
- New models are admin-gated and may be usage-priced
- Release-notes claims: confirm before you cite them
- Only as good as your data hygiene
Agents get an identity of their own
Agent identity entered public beta on July 15, and Glean’s release notes then carried an August 12, 2026 entry presenting it as a launch: “Let agents act as themselves with governed service credentials,” with “Agent Identity lets an agent use its own scoped service credentials instead of borrowing the identity of the person who invoked it.” Read the label precisely, because a rollout plan turns on it: the August 12 entry carries no “beta” qualifier, but it also carries no statement of general availability. So it has clearly moved past the July public-beta framing, and GA is unconfirmed — ask Glean directly before you write it into a control narrative. The substance is unchanged and still the most interesting governance idea in this set: access rights split from invocation rights, and audit trails that name the agent rather than whoever triggered it.
- Scoped service credentials per agent
- Past public beta as of August 12; GA unconfirmed
- Invocation rights split from access rights
- Audit trails name the agent
Ethical walls, enforced in the tool
Ethical Wall enforcement with Intapp reached general availability on July 23, syncing Intapp Walls policies into Threads, Vault, Review Tables and Shared Spaces. For a firm, conflict management is the gate that decides whether AI may touch matter data at all, so policy that lives only in the document management system is not enough. A July 23 expanded collaboration with Microsoft deepens Microsoft 365 integration, with Microsoft CELA adopting Harvey.
- Firm conflict policy carried into the AI
- Covers threads, vault and shared spaces
- Confirm coverage matches your walls
Beyond your own documents
The July 10 integration post is the clearest statement of scope: the corpus is no longer just the data room. Private files sit alongside public filings and licensed market data in a single view, which changes the diligence question from “can it read our documents” to “which third-party licenses do we already hold, and which does this tool require of us.”
- One view over three kinds of data
- License terms become part of the buy
- Source every number back to a filing
For this vertical tool, tell me: which underlying model it uses, exactly what proprietary data or workflow it adds on top, how it handles citations and source traceability, and where our data is stored and processed. Separate vendor marketing from verifiable fact.
Compare what this specialist does against what our existing Microsoft 365 Copilot and ChatGPT already cover. Identify the specific 20% of the workflow the specialist genuinely adds, and whether that 20% justifies the cost and integration.
Specialists that act, and specialists that must be right
Two high-stakes frontiers: customer-facing agents that take real actions for real users, and clinical tools where an error has a different weight entirely. Both show why domain guardrails and human accountability matter more than raw capability.
Sierra Agents that act
Began in customer experience — chat, voice, email and messaging, with outcome-based pricing and a reported 40% of the Fortune 50 as customers — and has now pushed past it. Horizon, announced July 16, targets goals that run for days, weeks or months, such as loan origination and prior authorization, rather than single conversations, and is confirmed as the current framing: two August posts describe agents that pursue business outcomes “over days, weeks or months.” A July 14 partnership makes SoftBank Corp. the exclusive sales partner in Japan, where the LINEMO deployment reports 97% resolution and 93% CSAT (vendor figures), expanding to the SoftBank and Y!mobile brands. Sierra acquired Takeoff on July 23, with the teams combining on Horizon. Founded by Bret Taylor; valued around $15.8B in 2026. The August window adds a new product this guide had not carried: Voice Personas, August 7, 2026. Around it sit “Rent the intelligence, own the relationship” (August 4), teaching agents to navigate janky IVR systems (August 11), defense in depth in the age of agents (August 13) and building in DACH (August 14), following “Agency: Secure, scalable sandboxes for agents” on July 29 and the Plaid partnership on August 3, which lets customers “securely connect their bank accounts with Plaid directly within Sierra’s agent.” The trajectory is consistent: past customer service, into voice and financial workflows.
- Voice Personas shipped August 7, 2026
- Lending and healthcare, not just support
- Priced on outcomes, not tokens or seats
- Horizon confirmed as the current framing in two August posts
- A months-long goal needs a named owner
Abridge Clinical documentation
Turns doctor–patient conversations into structured clinical notes in real time, integrated into Epic. In June 2026 NVIDIA collaborated with Abridge on a clinical-conversation foundation model. The scope is widening past documentation toward care delivery, payment and revenue cycle: the July 20 Altrina acquihire adds browser-based AI agents and EHR workflow automation, framed as a talent investment, and CHRISTUS Health expanded its deployment on July 21. Ambient documentation is still an accurate description, but increasingly names only a subset of the product. Between August 3 and August 19, 2026 Abridge published three posts, all customer or engineering content rather than corporate news: “Context-Aware Clinical Intelligence For Every Clinician at Partner Health Systems” (August 17), a clinician-builder piece (August 5) and a Deaconess Health post (August 4). Neither the Altrina acquihire nor the Northwestern Medicine win is restated in that window — no contradiction, but no reconfirmation either.
- Ambient notes are the beachhead, not the ceiling
- August posts are customer and engineering content, not product change
- Clinician reviews every note
- Consent and privacy are non-negotiable
OpenEvidence Clinical evidence
A medical answer engine grounded in licensed clinical literature, with multi-year content agreements with the NEJM Group and JAMA Network and enterprise deployments across major health systems.
- Evidence at the point of care
- Licensed, citable sources
- Clinical judgment stays human
A decision framework, not a vibe
The specialist-versus-generalist and buy-versus-build calls are recurring and expensive. Run them deliberately.
A $15B valuation is not a fit for your workflow
Specialist AI carries risks a general assistant does not: you are often paying a premium for someone else’s model, embedding sensitive data deeply, and betting on a market that is consolidating in real time — six deals across the eight vendors in this guide since July 10, 2026, the latest being Cursor’s own acquisition by SpaceX on August 14. The pace is uneven, not a metronome: five deals in seventeen days, then a fortnight with none, then the largest change of control in the set. Do not annualize a burst — and do not read a quiet fortnight as the end of one.
Rented intelligence
Many verticals wrap OpenAI or Anthropic. Confirm the domain layer is real before paying specialist prices for someone else’s model.
Overlap you already own
Your Copilot, ChatGPT or Claude may already cover most of the job. Buy the specialist only for the measured gap.
Data and compliance
Vertical tools touch your most sensitive data—legal, clinical, financial. Verify storage, processing, and HIPAA, SOC 2 or EU AI Act posture.
Lock-in
Proprietary data and workflow make these tools hard to leave. Check export and migration before you embed one in a critical path.
Consolidation, at speed
Six deals across these eight vendors since July 10, 2026. In the seventeen days to July 27: Cognition bought TierZero and welcomed The Interaction Company, Harvey bought Benchmark, Sierra bought Takeoff, Abridge took on the Altrina team. Then two quiet weeks. Then, on August 14, Cursor announced that it “has officially been acquired by SpaceX,” completing a process begun in April; no valuation or terms were disclosed. Harvey separately took a strategic investment from Goldman Sachs and J.P. Morgan on July 28, with no new valuation published. The tracked set now contains a vendor owned by a launch company, which sharpens the point rather than softening it: the specialist you sign with may not merely be a different company in a year, it may sit inside a group whose priorities have nothing to do with your codebase.
Do not infer ownership — but do not ignore the signal
This guide’s July 27 edition removed a draft claim that Cursor had been acquired: no source supported it and the date it gave was wrong. On August 14, 2026 Cursor announced the acquisition itself, first-party and dated. Both calls were right, and the arithmetic is the argument for the rule: the guide was wrong for zero days, and would have been wrong for eighteen. Keep the discipline and note its real price, which is low. A partnership, a joint model launch or a shared brand is evidence of a relationship, not of ownership. Get ownership, control and model supply written into the contract rather than inferred from a launch post — and put a change-of-control clause in anyway, because the inference does sometimes come true.
Valuation is not fit
A famous, well-funded tool can still be wrong for your one workflow. Evaluate on your job, not on the funding round.
| Before you buy a vertical tool | What to check | Who owns it |
|---|---|---|
| It wraps a frontier model | Which model, and whether the domain layer is genuinely added value | Technical evaluation |
| It touches sensitive data | Where data is stored and processed; compliance certifications | Security and legal |
| It overlaps tools you own | The exact gap versus your existing general assistant | Budget owner |
| It embeds deeply | Data export and migration path if you leave | Procurement |
| It produces work of record | Human review of every filing, diff, note or customer action | The professional of record |
| It may be acquired | Change-of-control terms, and what happens to price, roadmap and support | Procurement and the budget owner |
First-party evidence behind this guide
This is a fast-moving, fast-consolidating category, so these vendor pages are the most volatile sources in the whole set—re-verify before you act on any specific tool. Capability claims are anchored to each vendor’s own pages; valuations and funding are as reported and are noted as such in the text. Read one limit before you cite anything here: this guide lists roughly one representative dated article per vendor, and many other items referenced in the body were read on the vendors’ own channels without being separately cited below. Those include Devin Fusion on June 29, the Department of Energy memorandum of understanding on July 22, Cognition welcoming The Interaction Company on July 23, Sierra acquiring Takeoff the same day and shipping Voice Personas on August 7, Harvey’s Intapp Ethical Walls general availability on July 23, the Goldman Sachs and J.P. Morgan investment in Harvey on July 28, Harvey for Outlook on August 12, Cursor’s August releases from Google Workspace Plugins on August 3 to Git at any scale on August 18, the LTM partnership with Cognition on July 28, Hebbia’s August 4 post, and Abridge’s August customer posts. Glean is a deliberate exception twice over: its three July 29 items and its August 12 agent-identity entry appear only in Glean’s rolling release notes, which cannot anchor a dated claim, so they are stated in the body with that limitation named and no link. If you are relying on any one of these specifically, confirm it directly with the vendor.
AI Mindset